WVY-Cartoon-14m
A 14.7M-parameter cartoon character model trained from scratch on hand-drawn layers. You give it labels (skin, pose, outfit, hair, eyes, brows, nose, mouth, prop, background) and it draws the character, including combinations it was never shown. No pretrained anything. It was written in numpy and trained on 2 CPU cores.
βΆοΈ Try it in your browser
It's part of WVY, Starpower's research into generating intentionally: small models trained on data where every part is named, placed and swappable. The control comes from how the data is shaped, not from how much of it there is.
Every character above is a prompt the model never saw in training.
What it knows
132 hand-drawn pieces in 10 label groups, about 2.08 billion valid characters:
| Group | Options | Values |
|---|---|---|
| skin | 6 | grey, green, red-orange, rose, lilac, mint |
| pose | 6 | default (standing), arms up, waving, hands on hips, raise a cup, chubby |
| outfit | 17 + none | 7 original tees and long sleeves, no shirt, striped tees, hoodies (grey, black, pink), white button-up, flannel, jersey 23, puffer |
| hair | 11 | purple messy, spikes, black bowl, blond tuft, green dreads, buzz, mohawk, beanie, long hair + bangs, twists, space buns |
| eyes | 16 | sleepy, X, spirals, red dots, red blocks, big shiny, side-eye, hearts, stars, happy closed, crying, shades, heavy lids, dollar signs, dot eyes, angry |
| brows | 12 | soft, flat, thick blocks, slanted, one raised, heavy, angry, unibrow, worried, thin arches, slit brow, bushy |
| nose | 12 | n, u, long, pointy, hook, curly, dot, nostrils, triangle, clown, L, button |
| mouth | 14 | flat, squiggle, big smile, tongue out, frown, smirk, open, toothy grin, gold grill, shocked O, lips, buck teeth, drool, grimace |
| prop | 6 + none | cigar + smoke, double cup (right hand), liquor bottle (left hand), gold chain, headphones, raised double cup |
| background | 8 | indigo, purple, teal, white, rose, green, olive, charcoal |
Fit rules (what the training data always followed):
- Shirts only go on the standing (default) body.
- Hand-held props (double cup, liquor bottle) only go on the standing body.
- The raised cup only goes with the "raise a cup" pose.
How it works
labels β 110 on/off switches (one per trait value)
β 768 neurons β 512 neurons β 96 Γ 96 Γ 3 pixels
- Each switch is a learned recipe for one part. Turning on "heart eyes" adds that recipe. The later layers work out placement and what covers what (hair over brows, a cup over a hand), then paint.
- Changing one label changes one part. Swap one trait on a never-seen character and 94β99.8% of the pixel change lands inside that part's own area.
- Size follows output detail, not data. 96.7% of the parameters are the final painting layer. Doubling the combinations adds 0 parameters, and a new part adds 768.
| Layer | Parameters | Share |
|---|---|---|
| 110 switches β 768 | 85,248 | 0.6% |
| 768 β 512 | 393,728 | 2.7% |
| 512 β 96Γ96Γ3 | 14,183,424 | 96.7% |
| Total | 14,662,400 |
Training
- Art: 55 pieces hand-drawn in Adobe Fresco, plus 77 new pieces drawn to match the style (new poses and body shapes in all 6 skins, outfits, hair, eyes, mouths, props). The artist reviewed and corrected the new pieces.
- Data: 10,000 characters stacked from the layers with exact labels, at 192 px, downscaled to 96 px for training.
- Held out: 300 random combinations, plus 3 trait pairs that were never shown together: mohawk + heart eyes, pink hoodie + red-orange skin, and waving + purple messy hair.
- Pretrain: 8,000 steps on all 10k (about 35 minutes on 2 CPU cores).
- Fine-tune: 2,000 steps on 3,138 "good ones" (images with at least 3 favorite traits). These are the default weights,
wvy_cartoon_14m.npz. The pretrained weights are inwvy_cartoon_14m_pretrained.npz.
Benchmark (420 prompts it never saw)
For each test prompt and each trait group, every other valid option is swapped in and re-rendered by hand. The check is whether the model's image sits closest to the correct one. The comparison is a closest-label lookup that returns the training image sharing the most traits, which is what pure memorizing would get.
| System | Every trait right | Traits right | PSNR |
|---|---|---|---|
| WVY-Cartoon-14m (pretrained) | 99.5% | 99.95% | 35.8 dB |
| WVY-Cartoon-14m (fine-tuned, default) | 99.3% | 99.92% | 35.5 dB |
| Closest-label lookup | 0% | 56.6% | 15.5 dB |
All three never-together pairs scored 100%, except waving + purple messy hair at 95%. Those misses came from test prompts that broke a fit rule. It makes about 2,500 images a second on 2 CPU cores.
Fine-tuning is a trade: on never-seen prompts in the good-ones style, error dropped 6%. On everything else it rose 15%.
Limitations
Left to right: a hoodie on the arms-up pose (never seen, comes out smeared) Β· a raised cup while waving (comes out see-through) Β· 50% hearts + 50% shades (a double exposure, not a new eye) Β· "no hair" (never an option, so a bald blur).
- 96Γ96 pixels. Faces are readable but soft.
- One image per prompt. The same labels always give the same picture, with no variety.
- Closed vocabulary. It can only draw the 132 pieces it was given.
- Faint ghosts of other parts can show, even when every trait is right.
- It learned the fit rules, not the reasons. Combos outside the rules come out half-blended.
- The benchmark measures faithfulness to the drawings, not taste.
Use it
Needs only numpy and Pillow.
python3 wvy_cartoon.py --list # every label
python3 wvy_cartoon.py --random -o me.png # a random valid character
python3 wvy_cartoon.py --skin lilac --pose "raise a cup" --hair "space buns" \
--eyes hearts --mouth "gold grill" --background teal -o cup.png
Any trait you leave out is filled in at random, within the fit rules.
Files
| File | What it is |
|---|---|
wvy_cartoon_14m.npz |
Fine-tuned weights (default), float16 |
wvy_cartoon_14m_pretrained.npz |
Pretrained weights, float16 |
vocab.json |
Label groups and values, in input order |
wvy_cartoon.py |
Labels β image, enforcing the fit rules |
benchmark.json |
Full benchmark results |
training/ |
The scripts used to build the dataset, train, benchmark and pick outputs (paths point to the original workspace) |
media/ |
Explainer video and figures |
Draw the parts. Let it learn the grammar. Β· WVY Β· Starpower